TencentDB saves your progress

Article by Nacata Security, 04/08/2026

How many times have you watched an AI agent start from scratch explaining the same things?

TencentDB Agent Memory is here to put an end to that endless repetition.

Every time you start a new session, you end up re-explaining the project context and decisions already made. The cost of that restart accumulates in silence.

The problem wasn’t the agent’s memory: it was that no one had built a system to let team knowledge survive across sessions and frameworks.

TencentDB Agent Memory turns that knowledge into reusable assets.

The tool organizes what a team learns into four assets: Chat Memory, which stores valuable interactions; Skills, proven workflows; LLM-Wiki, structured documentation in navigable pages; and CodeGraph, a code index that maps which functions affect which others. Each asset is recorded, versioned, and assigned with specific permissions to the agents that need it.

This isn’t storage: it’s a saved game.

When a new agent comes on board, it doesn’t start from page one. It loads the team’s save file and inherits the accumulated context. Assets have controlled visibility: Chat Memory and Skills are private by default, and sharing them is an explicit action. Each agent receives only the assets it needs for its role, with no noise or unnecessary exposure.

A review agent already knows which modules must not be touched.

Team memory is no longer lost.

TencentDB Agent Memory is available on GitHub in its beta v2 release and evolving fast. Installation spins up all three services—memory-core, memory-hub, and proxy—in a single command. What comes next changes how agents learn within a team.

But there is a technical layer worth understanding.

Conversations pass through an asynchronous pipeline that refines them into granularity levels, from L0 to L3. In normal retrieval, L2 and L3 provide fast context; when a specific piece of data is needed, the system combines BM25, vector retrieval, and RRF to drill down to L0. Item limits and character budgets prevent memory from saturating the agent’s context window.

The system decides how much memory to inject, not the user.

That means you don’t have to manually manage which context is passed to each agent. TencentDB Agent Memory filters first by team, user, and agent permissions, then retrieves only what is relevant to the active query. Less noise, more precision.

The benchmark is also on the table.

The tool includes evaluations using PersonaMem, which measures whether an agent correctly applies user information after long interactions. The project acknowledges that memory standards for agents are not yet consolidated and opens the door to contributions: bug reports, new framework adapters, and benchmark reproductions.

Open source, open to anyone who wants to build.

The repository is on GitHub under TencentCloud and accepts pull requests, feature suggestions, and documentation fixes. The community decides where it grows.

How much context has your team already lost between sessions?

Security is not improvised, it is audited. At Nacata Security we detect vulnerabilities and protect your company, because a single flaw can cost you everything you have built.

Related articles

Nacata Security, reach out to us anytime